Research map / Architectures and efficiency / Quantization
MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling
Yu Zhang, Hui-Ling Zhen, Mingxuan Yuan, Bei Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| FP8 training with MOSS instead of BF16 | Qwen-3-14B/32B · MATH500, GPQA, C-Eval, AIME24, MMLU-Redux | accuracy: 96.8 -> 96.7 (14B MATH500) | No clear effect |
| use two-level microscaling instead of per-group quantization (ablation) | MOSS · activation SNR (LLaMA-2-7B) | SNR (dB): 34.5 -> 37.5 early; 32.9 -> 36.0 late | Improved |
| use two-level microscaling instead of per-tensor quantization (ablation) | MOSS · activation SNR (LLaMA-2-7B) | SNR (dB): 28.3 -> 37.5 early | Improved |
| use MOSS kernel instead of COAT per-group GEMM (ablation) | MOSS · FP8 GEMM on H800 | avg runtime (ms): 3.73 -> 0.77 | Improved |
| MOSS GEMM vs DeepGEMM | MOSS · FP8 GEMM on H800 | avg runtime (ms): 0.54 -> 0.77 | Hurt |
| use automatic scaling instead of just-in-time scaling (ablation) | MOSS · LLaMA-2-7B fine-tuning | throughput (tokens/s): 38,642 -> 41,998 (1.087x) | Improved |
| use automatic scaling instead of just-in-time scaling (ablation) | MOSS · LLaMA-2-7B MAmmoTH | GSM8K accuracy: 64.9% -> 64.7% | No clear effect |
| increase rescaling interval from 500 to 2000 (ablation) | MOSS · LLaMA-2-7B NumGLUE | accuracy: 59.4% -> 58.1% | Hurt |
| use FP8 MOSS instead of COAT for memory | MOSS · peak activation memory | GB: 28.6 -> 23.5 | Improved |
| use MOSS FP8 instead of BF16 for pretraining throughput | MOSS · OLMo-7B | training speedup: 1.34x vs BF16 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- COAT — COAT: Compressing Optimizer states and Activations for Memory-Efficient FP8 Training (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Fully FP8 GEMM LLM Training at Scale (NeurIPS 2025 · similarity 0.91)
- µnit Scaling: Simple and Scalable FP8 LLM Training (ICML 2025 · similarity 0.91)
- Scaling FP8 training to trillion-token LLMs (ICLR 2025 · similarity 0.90)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.90)
- COAT: Compressing Optimizer states and Activations for Memory-Efficient FP8 Training (ICLR 2025 · similarity 0.88)
- Quartet: Native FP4 Training Can Be Optimal for Large Language Models (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
